Multi-class classification of construction hazards via cognitive states assessment using wearable EEG. (August 2022)

• Record Type: Journal Article
• Authors: Jeon, JungHo; Cai, Hubo
• Is Part Of: Advanced engineering informatics. Volume 53(2022)
• Publication Date: 2022-08
• Subjects: Construction safety -- Hazard classification -- Electroencephalogram (EEG) -- Wearable EEG -- Virtual reality (VR) -- EEG classifier; Computer-aided engineering -- Periodicals; Engineering -- Data processing -- Periodicals; 620.00285
• DOI: 10.1016/j.aei.2022.101646
• Languages: English
• ISSNs: 1474-0346

Abstract

Abstract: Improving workers' safety and health is one of the most critical issues in the construction industry. Research attempts have been made to better identify construction hazards on a jobsite by analyzing workers' physical responses (e.g., stride and balance) or physiological responses (e.g., brain waves and heart rate) collected from the wearable devices. Among them, electroencephalogram (EEG) holds unique potential since it reveals abnormal patterns immediately when a hazard is perceived and recognized. Unfortunately, the unproven capacity of EEG signals for multi-hazard classification is a primary barrier towards ubiquitous hazard identification in real-time on jobsites. This study correlates EEG signal patterns with construction hazard types and develops an EEG classifier based on the experiments conducted in an immersive virtual reality (VR) environment. Hazards of different types (e.g., fall and slip/trip) were simulated in a VR environment. EEG signals were collected from subjects who wore both wearable EEG and VR devices during the experimentation. Two types of EEG features (time-domain/frequency-domain features and cognitive features) were extracted for training and testing. A total of eighteen advanced machine learning algorithms were used to develop the EEG classifier. The initial results showed that the LightGBM classifier achieved 70.1% accuracy based on the cognitive feature set for the 7-class classification. To improve the performance, the input data was relabeled, and three strategies were designed and tested. As a result, the combined approach (two-step ensemble classification) achieved 82.3% accuracy. As such, this study not only demonstrates the feasibility of coupling wearable EEG, VR, and machine learning to differentiate jobsite hazards but also provides strategies to improve multi-class classification performance. The research results support ubiquitous hazard identification and thereby contribute to the safety of the construction workplace.